Apollo Economist: No AI Payoff Outside Tech — What the Claim Says and What It Does Not

Apollo chief economist Torsten Slok published a short note on October 24, 2026, titled "No Signs of AI Payoff Outside the Tech Sector," and it climbed past 300 points on Hacker News within hours. His claim: "the technology has delivered zero financial payoff to corporate America outside of tech." Revenue per worker and profit margins are rising for the Mag 7 but flat for the S&P 493, and there is no sign of an AI lift in productivity growth.

The note is five charts and a handful of paragraphs. That makes it easy to quote and easy to over-read. This post lays out exactly what Apollo says, which parts are data and which are opinion, what the Hacker News thread added, and what it means if you build or buy AI systems and your budget is about to be reviewed. It fits alongside our earlier coverage of Gartner's finding that 13 percent of CIOs report significant AI value and the McKinsey State of AI ROI read.

Quick reference: what the note says

What exactly does Apollo claim?

Slok makes four separate statements, and they carry different weight.

First, revenue per worker is rising for the Mag 7 but flat for the S&P 493 (inflation adjusted to 2019 Q4, per the chart note). Second, profit margins show the same split. Third, there are "still no signs of AI boosting productivity growth," illustrated with the utilization-adjusted total factor productivity series from the San Francisco Fed. Fourth, and this is the part that travelled furthest, he says "mainstream corporate America is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins."

The first three are chart-backed observations about aggregates. The fourth is an interpretation: the note's last two charts draw on Ramp's AI Index and Ramp chief economist Ara Kharazian's work, which track business spending on AI from card and bill-pay data. We could not see the underlying series beyond what the page shows, so we cannot confirm exactly which spending cuts "pulling back" refers to. Treat that sentence as Apollo's reading of the Ramp data, not as a measured fact about all large companies.

What the note does not show

Three gaps matter before you quote it.

Aggregates hide timing. Revenue per worker across 493 firms is a blunt instrument. Many of the companies in that group spent 2025 and early 2026 on pilots, data cleanup and security reviews. Flat margins today are what you would expect from firms that have only just paid the cost of adoption. The note does not claim otherwise; its own closing paragraph asks whether companies "can turn increasingly cheap intelligence into higher revenue growth."

Correlation with the Mag 7 is not a control group. The Mag 7 were already the most profitable and most software-heavy companies. Their margin expansion has many drivers, and the note does not isolate AI. The same caveat cuts the other way: flat S&P 493 numbers do not prove AI has no effect, because other forces, such as tariffs, wages and interest rates, move the same lines.

"Zero" is a strong word. The note uses it about aggregate financial payoff. It does not examine individual firms, functions or use cases. Our own coverage of measured workplace studies shows real gains in narrow tasks that fail to reach the bottom line because review and rework absorb them.

Why the "cost center" framing matters to builders

Slok's most useful sentence for practitioners is that "mainstream corporate America does not need the latest model and is unwilling to pay for it." If that holds, it changes what sells. Frontier-model pricing is built around buyers who treat tokens as growth spend, such as VC-backed startups and coding-agent heavy teams. A second tier of buyers wants good-enough models at lower prices, with predictable bills.

We saw the other half of this in August, when Ramp's data showed the top 1 percent of companies concentrating AI spend per employee. If spend is concentrated in a few firms and stalls in the rest, the macro effect Slok describes follows almost by arithmetic: the note says as much when it argues that effects stay "very limited" as long as agentic AI is confined to a small part of the economy.

It also lines up with the cost of checking AI output. Surveys we reviewed in the AI verification tax found that time saved drafting is partly spent reviewing and redoing. That is a plausible mechanism for flat margins, though the Apollo note does not test it.

What developers on Hacker News said

The Hacker News thread had roughly 300 comments at the time of writing. Reading it, the objections fell into a few groups. These are commenters' views, not verified facts.

• "Too early." User jstanley quoted the 1987 line about seeing the computer age everywhere except the productivity statistics, and briHass argued that the big model improvements are only about nine months old and business change is slow.
• Electrification analogy. secabeen argued that tech is the most software-defined sector, so it moves first, while other industries need capital changes. They compared it to factories built around water wheels that only gained when rebuilt around electric motors.
• Validation gap. tapoxi wrote that software teams can use a non-deterministic system because test suites act as validation gates, and other industries lack that, so checking AI output by hand can cost more than doing the work.
• Not even in tech. realusername reported that productivity return in their day job has been slightly negative so far, and bob1029 said payoff is hard to achieve even inside the sector, with development teams the biggest problem.
• Whose baseline? drivebyhooting pointed out that a similar sentence could have been written about the past 15 years of tech and the wider economy.
• Budget season. cmiles8 predicted that companies are cutting AI budgets and internal AI lab teams as corporate planning begins. That is one commenter's observation, not data.

The thread's strongest consensus was not that Slok is wrong or right, but that aggregate data is a lagging instrument for a technology whose adoption is uneven.

The productivity-paradox precedent

Economists have seen this shape before. Robert Solow's remark about computers and productivity statistics became shorthand for a gap between visible technology adoption and measured output. Later work attributed much of the delay to complementary investment: firms had to reorganize processes before the gains appeared. Our piece on the AI value and GDP measurement problem covers why some of the value may also sit outside the statistics, for example in free consumer use and quality improvements.

The same logic shows up in field research. A Varick analysis we covered argued applied AI fails when it is bolted onto an unchanged process. If that is right, the S&P 493 flatline reflects process redesign that has not happened yet, and Slok's real question, whether firms can convert cheaper intelligence into revenue, is about organization, not model quality.

What this means for what you build or pay

Finance commentary only matters to this blog when it changes decisions, so here is the practical reading.

1. Expect scrutiny of AI line items. If boards read notes like this, token spend gets reviewed as a cost. Prepare a one-page answer per workflow: what it replaces, what it costs per task, and how the result is checked.
2. Right-size the model. Slok's claim that mainstream buyers do not want the latest model is a prompt to test smaller or older models on your actual tasks. Many routine workloads do not need a frontier tier.
3. Count verification time. A workflow that saves 30 minutes of drafting but adds 25 of review is a weak case. Measure both sides.
4. Tie usage to a business metric. Seats and prompts are activity, not payoff. Pick revenue per rep, tickets closed, cycle time or defects, and report that.
5. Pick process-first projects. Where gains are reported, they usually come after a workflow was redesigned around the tool rather than adding the tool to the old workflow.

How to check the claim yourself

You do not need to take either side on trust. Pull the San Francisco Fed's TFP series and the Ramp AI Index pages linked above and see whether the lines match the charts. Compare revenue per employee for an equal-weight S&P 500 index against the cap-weighted one, and look at a single industry you know. If you work inside a company, run the same test on your own AI projects: baseline metric, intervention, six-month comparison.

Bottom line

Apollo's note is a credible economist's read of aggregate data and a useful warning about how finance teams may start to view token budgets. It is not proof that AI cannot pay off outside tech, and it says itself that the question is open. The practical response is the same under either reading: measure outcomes, control cost, and redesign the process around the tool.

Figures and quotes are from the Apollo note and linked sources as of October 24, 2026; the underlying data may be revised.

Related reading

• Gartner: 13 percent of CIOs report significant AI value
• McKinsey State of AI 2026: ROI and agentic coding
• Top 1 percent of companies and AI spend per employee (Ramp)
• The AI verification tax: rework, surveys and how to cut it
• AI value and the GDP measurement problem
• Why applied AI does not work without process redesign
• AI coding agents: more code, not more software